Image Processing-Driven Spectrum Sensing with Small Training Samples

To resolve the problems of high computational complexity in strong noise environment, infeasibility of gaining large number of labeled samples and low detection probability, an Image Denoising and Classification driven Spectrum Sensing (IDCSS) method is proposed. Firstly, time-frequency transformation is employed to convert radio numerical signals into images. Then, as received signals of cognitive users and noise are highly correlated under strong noise environments, a novel Generative Adversarial Network (GAN) is designed to enhance the number and quality of samples of cognitive user signals. In the generator, residual-long-short-term memory network is designed to replace U-Net skip connection, realizing denoising and multi-scale features extraction. Loss function based on entropy is designed to optimize robustness to noise. A multi-dimensional discriminator is designed to enhance the quality of the generated image and retain the image details of the low signal-to-noise ratio cognitive user signals. Finally, the generated high-quality samples are used as labeled data, and the real samples combine to train the classifier to realize the recognition and classification of the spectrum occupancy state. Simulation results show that the proposed algorithm has better detection performance by comparing it with the state-of-the-art methods.